[llvm] [AArch64] Use SVE rev for full reverse shuffles with SVE128 (PR #224589)

[David Green via llvm-commits]

================
@@ -16195,11 +16195,23 @@ SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
                     DAG.getNode(AArch64ISD::NVCAST, DL, BSVT, V1)));
   }
 
-  if (((NumElts == 8 && EltSize == 16) || (NumElts == 16 && EltSize == 8)) &&
+  if (((NumElts == 8 && EltSize == 16) || (NumElts == 16 && EltSize == 8) ||
+       (NumElts == 4 && EltSize == 32)) &&
       ShuffleVectorInst::isReverseMask(ShuffleMask, ShuffleMask.size())) {
-    SDValue Rev = DAG.getNode(AArch64ISD::REV64, DL, VT, V1);
-    return DAG.getNode(AArch64ISD::EXT, DL, VT, Rev, Rev,
-                       DAG.getConstant(8, DL, MVT::i32));
+    // For sve128 we can use a REV full vector reverse.
+    if (Subtarget->isSVEorStreamingSVEAvailable() &&
+        Subtarget->getSVEVectorSizeInBits() == 128) {
+      EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
+      V1 = convertToScalableVector(DAG, ContainerVT, V1);
+      SDValue Rev = DAG.getNode(ISD::VECTOR_REVERSE, DL, ContainerVT, V1);
+      return convertFromScalableVector(DAG, VT, Rev);
+    }
+
+    if (EltSize != 32) {
----------------
davemgreen wrote:

Yep - perfect shuffles will handle any 4 element masks. It seems OK to handle these here though, if we have already checked the mask is a reverse.

https://github.com/llvm/llvm-project/pull/224589